Papers by Sharan Narang
Compute Optimal Scaling of Skills: Knowledge vs Reasoning (2025.findings-acl)
Copied to clipboard
| Challenge: | Scaling laws are a critical component of the LLM development pipeline, but little is known about whether the COs of individual skills such as mathematical reasoning, question answering (QA) or coding, align with these APEs. |
| Approach: | They examine knowledge-based QA and code generation to find out whether skill-dependent scaling is an artefact of the pretraining datamix. |
| Outcome: | The proposed scaling laws are skill-dependent, and knowledge and code exhibit fundamental differences in scaling behaviour when corrected for datamix differences. |
Scaling Laws vs Model Architectures: How does Inductive Bias Influence Scaling? (2023.findings-emnlp)
Copied to clipboard
Yi Tay, Mostafa Dehghani, Samira Abnar, Hyung Chung, William Fedus, Jinfeng Rao, Sharan Narang, Vinh Tran, Dani Yogatama, Donald Metzler
| Challenge: | Existing studies on the scaling properties of model architectures have not explored the impact of inductive biases on scaling behaviour. |
| Approach: | They conduct extensive experiments to understand scaling behaviour of ten different model architectures. |
| Outcome: | The results show that the best performing model can fluctuate at different scales. |
Do Transformer Modifications Transfer Across Implementations and Applications? (2021.emnlp-main)
Copied to clipboard
Sharan Narang, Hyung Won Chung, Yi Tay, Liam Fedus, Thibault Fevry, Michael Matena, Karishma Malkan, Noah Fiedel, Noam Shazeer, Zhenzhong Lan, Yanqi Zhou, Wei Li, Nan Ding, Jake Marcus, Adam Roberts, Colin Raffel
| Challenge: | Currently, the Transformer is the de facto architecture of choice for processing sequential data. |
| Approach: | They evaluate the Transformer architecture and its modifications in a shared experimental setting . they conjecture that performance improvements may strongly depend on implementation details . |
| Outcome: | The proposed improvements do not significantly improve performance, the authors find . the proposed improvements are either developed in the same codebase or are minor changes . |
Character-Aware Models Improve Visual Text Rendering (2023.acl-long)
Copied to clipboard
Rosanne Liu, Dan Garrette, Chitwan Saharia, William Chan, Adam Roberts, Sharan Narang, Irina Blok, Rj Mical, Mohammad Norouzi, Noah Constant
| Challenge: | Current image generation models struggle to produce well-formed visual text due to lack of character-level input features. |
| Approach: | They conduct a series of experiments to compare character-aware vs. character-blind text encoders to determine their spelling ability. |
| Outcome: | The character-aware models outperform character-blind models on a range of novel text rendering tasks. |
Effective Long-Context Scaling of Foundation Models (2024.naacl-long)
Copied to clipboard
Wenhan Xiong, Jingyu Liu, Igor Molybog, Hejia Zhang, Prajjwal Bhargava, Rui Hou, Louis Martin, Rashi Rungta, Karthik Abinav Sankararaman, Barlas Oguz, Madian Khabsa, Han Fang, Yashar Mehdad, Sharan Narang, Kshitiz Malik, Angela Fan, Shruti Bhosale, Sergey Edunov, Mike Lewis, Sinong Wang, Hao Ma
| Challenge: | Large language models (LLMs) are rapidly deployed and continue to evolve through scaling. |
| Approach: | They propose a method to train strong long-context LLMs that are capable of utilizing massive context windows of up to 32,000 tokens. |
| Outcome: | The proposed model can surpass gpt-3.5-turbo-16k's overall performance on long-context benchmarks with a cost-effective instruction tuning procedure that is free of expensive annotations. |
ByT5: Towards a Token-Free Future with Pre-trained Byte-to-Byte Models (2022.tacl-1)
Copied to clipboard
Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel
| Challenge: | a number of pre-trained language models use sequences of tokens corresponding to word units . token-free models that operate directly on raw text have many advantages . |
| Approach: | They propose a standard Transformer architecture that can be used to process byte sequences . they also characterize trade-offs in terms of parameter count, training FLOPs, and inference speed . |
| Outcome: | The proposed model is more robust to noise and more robust on spelling and pronunciation tasks. |
Understanding HTML with Large Language Models (2023.findings-emnlp)
Copied to clipboard
Izzeddin Gur, Ofir Nachum, Yingjie Miao, Mustafa Safdari, Austin Huang, Aakanksha Chowdhery, Sharan Narang, Noah Fiedel, Aleksandra Faust
| Challenge: | Large language models have shown exceptional performance on a variety of natural language tasks, but their capabilities for HTML understanding have not been fully explored. |
| Approach: | They propose to use HTML understanding models to parse HTML and perform HTML navigation tasks with a large-scale HTML dataset. |
| Outcome: | The proposed models perform 50% more tasks with 192x less data than the previous best supervised model. |